5. Discussion
Analysis of longitudinal data is a crucial statistical approach that is widely used
in the health, social, and biological sciences. In this article we discussed four
statistical models for longitudinal data analysis-ANOVA, MANOVA, MRM, and
GEE. ANOVA and MANOVA are well-known and easy to manipulate in SAS,
and both models assume interval measurement and normally distributed errors that
are homogeneous across groups. ANOVA assumes compound symmetry which has
little validity for longitudinal. And, MAOVA does not permit missing data. They
only estimate and compare the group mean and not informative about individual
growth. MRM models are quite widely used for analysis of longitudinal data.
These models can be applied to ordinal outcomes or nominal or count outcomes
that have a Poisson distribution, which we have not discussed in this paper. The
advantage of MRM is that missing data are ignorable if the missing responses can
be explained either by covariates in the model or by the available responses from a
given subject. The disadvantage is that full-likelihood methods are more
computationally complex than quasi-likelihood methods such as GEE. When the
scientific interest is in estimation and inference of the regression parameters and
not of the variance-covariance structure, GEE provides standard errors that are
robust to mis-specifiction of the variance-covariance structure. Also, as stated,
GEE is often used as a general and computationally convenient method. In fact,
software for performing GEE analysis is available in most of the major statistical
software packages. The disadvantage is that missing data are only ignorable if the
missing data are explained by covariates in the model. This is a more stringent
assumption than MRM, and therefore GEE models have somewhat limited
applicability to incomplete longitudinal data.